RabotniKuma/Fast-Math-R1-14B
RabotniKuma/Fast-Math-R1-14B is a 14.8 billion parameter language model developed by Hiroshi Yoshihara, Yuichi Inoue, and Taiki Yamaguchi, based on the DeepSeek-R1-Distill-Qwen-14B architecture. It is specifically optimized for mathematical reasoning tasks, achieving up to 30% faster inference while maintaining accuracy compared to its base model. This model excels at solving difficult math problems, as demonstrated by its performance in the Kaggle AI Mathematical Olympiad.
Loading preview...
Overview
RabotniKuma/Fast-Math-R1-14B is a 14.8 billion parameter model developed by Hiroshi Yoshihara, Yuichi Inoue, and Taiki Yamaguchi, presented in the paper "A Practical Two-Stage Recipe for Mathematical LLMs." This model significantly enhances the performance of DeepSeek-R1-Distill-Qwen-14B for mathematical reasoning, achieving up to 60% (average 30%) faster inference while maintaining accuracy. It was a 9th place solution in the Kaggle AI Mathematical Olympiad - Progress Prize 2.
Key Capabilities
- Optimized Mathematical Reasoning: Achieves high accuracy on complex math problems, specifically AIME 2024 and 2025 benchmarks.
- Efficiency: Demonstrates reduced mean output tokens and faster inference times compared to its base model across various token budgets.
- Advanced Training Methodology: Utilizes a two-stage training approach involving Supervised Fine-Tuning (SFT) on a high-difficulty dataset and Generative Reinforcement Learning with Policy Optimization (GRPO) for efficient reasoning.
- Token Efficiency: GRPO training incorporates format, cosine, and length-based rewards to discourage overthinking and promote concise, accurate outputs.
Good For
- Solving Difficult Math Problems: Particularly suited for competitive mathematics and complex algebraic or geometric challenges.
- Applications Requiring Efficient Mathematical Inference: Ideal for scenarios where both accuracy and speed are critical for mathematical problem-solving.
- Research in Mathematical LLMs: Provides a strong baseline and methodology for further exploration into efficient mathematical reasoning models.